CN-122022062-A - Water quality prediction method, device, equipment and medium integrating uncertainty quantification

CN122022062ACN 122022062 ACN122022062 ACN 122022062ACN-122022062-A

Abstract

The invention relates to the technical field of water environment monitoring, and discloses a water quality prediction method, a device, equipment and a medium integrating uncertainty quantification, which comprehensively utilize the differentiation capability of each model in the aspects of capturing time sequence dependence, local characteristic and long-range association by integrating a plurality of network models, the method effectively overcomes the structural deviation and the contingency of a single model, enables the final prediction result to be more accurate and stable, reduces the excessive dependence of the prediction result on a specific model or data characteristics, and improves the overall robustness of the method. And then, introducing a joint probability distribution model, so that the conditional probability distribution of the real water quality value under the given multi-model prediction information, namely a prediction interval and a final prediction reference value, is obtained, and the span from deterministic point prediction to probabilistic interval prediction is realized.

Inventors

ZHANG CHI
YIN WEI
GAO YONG
LIU SHENG
ZHOU YAN
LIU YONG
GAO JUNKAI
HAO RUI

Assignees

中国长江三峡集团有限公司

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (10)

1. A method of predicting water quality with fusion uncertainty quantization, the method comprising: Acquiring an environment driving factor value at a future time, inputting the environment driving factor value into a plurality of pre-constructed water quality prediction models to obtain the future time water quality prediction value output by each water quality prediction model, wherein the pre-constructed water quality prediction models are respectively obtained by training a plurality of network models through a historical data set, the historical data set comprises a historical water quality observation sequence and a historical environment driving factor sequence, the historical water quality observation sequence comprises water quality observation values at a plurality of historical times, and the historical environment driving factor sequence comprises environment driving factor values at a plurality of historical times; and generating a prediction interval and a final prediction reference value of the water quality at the future time according to the water quality predicted value at the future time output by each water quality prediction model and a pre-constructed joint probability distribution model, wherein the joint probability distribution model is used for describing the statistical dependency relationship between the predicted value and the observed value.
2. The method for predicting water quality by fusion uncertainty quantization according to claim 1, wherein the construction process of each water quality prediction model is as follows: And training a plurality of network models by taking the environmental driving factor value in the historical data set as a characteristic and the water quality observation value as a label to obtain a plurality of water quality prediction models, wherein the plurality of network models comprise a long-short-term memory network model, a time sequence convolution network model and a Informer network model.
3. The method for predicting water quality by fusion uncertainty quantization according to claim 1, wherein the step of generating a prediction interval and a final prediction reference value of water quality at a future time according to the future time water quality prediction value output by each water quality prediction model and a pre-constructed joint probability distribution model comprises the steps of: converting the future time water quality predicted value output by each water quality predicted model into a cumulative probability value obeying uniform distribution; combining the cumulative probability value of the water quality predicted value at the future moment with a pre-constructed joint probability distribution model to calculate a conditional cumulative distribution function of the water quality true value at the future moment; and determining a prediction interval under a specified confidence level according to a conditional accumulation distribution function of the real value of the water quality at the future moment, and selecting a predicted value of the key quantile point as a final predicted reference value of the water quality at the future moment.
4. A method of predicting water quality with integrated uncertainty quantization as set forth in claim 3, wherein said step of converting future time water quality predictions output by each of said water quality prediction models into cumulative probability values subject to uniform distribution comprises: Respectively inputting the historical environment driving factor sequences into the water quality prediction models to obtain historical water quality prediction sequences output by the water quality prediction models, wherein the historical water quality prediction sequences comprise water quality prediction values at a plurality of historical moments; Respectively determining the historical water quality observation sequence and the edge probability distribution model of the historical water quality prediction sequence output by each water quality prediction model in a fitting mode; and converting the future time water quality predicted value output by each water quality prediction model into a cumulative probability value obeying uniform distribution through an edge cumulative distribution function corresponding to each edge probability distribution model.
5. The method according to claim 4, wherein the step of determining the edge probability distribution model of the historical water quality observation sequence and the historical water quality prediction sequence output by each water quality prediction model by fitting respectively includes: acquiring a plurality of candidate edge probability distribution types; Fitting the historical water quality observation sequence and the historical water quality prediction sequence output by the water quality prediction model by adopting each edge probability distribution type to obtain a plurality of candidate edge probability distribution models corresponding to each sequence; and selecting an optimal candidate edge probability distribution model corresponding to each sequence as an edge probability distribution model according to a plurality of candidate edge probability distribution models corresponding to each sequence.
6. The method for predicting water quality by fusion uncertainty quantization according to claim 1, wherein the method for constructing the joint probability distribution model is as follows: Respectively inputting the historical environment driving factor sequences into the water quality prediction models to obtain historical water quality prediction sequences output by the water quality prediction models, wherein the historical water quality prediction sequences comprise water quality prediction values at a plurality of historical moments; and constructing a joint probability distribution model between the water quality observation value and the water quality prediction value output by each water quality prediction model according to the historical water quality observation sequence and the historical water quality prediction sequence output by each water quality prediction model.
7. The method according to claim 6, wherein the step of constructing a joint probability distribution model between the water quality observation value and the water quality prediction value output by each of the water quality prediction models based on the historical water quality observation sequence and the historical water quality prediction sequence output by each of the water quality prediction models, comprises: Respectively determining the historical water quality observation sequence and the edge probability distribution model of the historical water quality prediction sequence output by each water quality prediction model in a fitting mode; Converting the values in each sequence into cumulative probability values obeying uniform distribution through the edge cumulative distribution function corresponding to each edge probability distribution model; And carrying out parameter estimation on the Vine Copula structure based on the cumulative probability value, and constructing a joint probability distribution model for representing the dependency relationship between the water quality observation value and the water quality prediction value output by each water quality prediction model.
8. A water quality prediction apparatus incorporating uncertainty quantization, the apparatus comprising: The system comprises a future water quality prediction module, a prediction module and a prediction module, wherein the future water quality prediction module is used for acquiring environmental driving factor values at future time and inputting the environmental driving factor values into a plurality of pre-constructed water quality prediction models to obtain future time water quality predicted values output by the water quality prediction models, the plurality of pre-constructed water quality prediction models are respectively obtained by training a plurality of network models through a historical data set, the historical data set comprises a historical water quality observation sequence and a historical environmental driving factor sequence, the historical water quality observation sequence comprises water quality observation values at a plurality of historical time, and the historical environmental driving factor sequence comprises environmental driving factor values at a plurality of historical time; And the final prediction module is used for generating a prediction interval and a final future time water quality prediction reference value according to the future time water quality prediction value output by each water quality prediction model and a pre-constructed joint probability distribution model, and the joint probability distribution model is used for describing the statistical dependency relationship between the prediction value and the observation value.
9. An electronic device, comprising: A memory and a processor in communication with each other, the memory having stored therein computer instructions, the processor executing the computer instructions to perform the fusion uncertainty quantified water quality prediction method of any one of claims 1-7.
10. A computer-readable storage medium having stored thereon computer instructions for causing a computer to perform the fusion uncertainty quantified water quality prediction method of any one of claims 1-7.

Description

Water quality prediction method, device, equipment and medium integrating uncertainty quantification Technical Field The invention relates to the technical field of water environment monitoring, in particular to a water quality prediction method, device, equipment and medium integrating uncertainty quantification. Background The method accurately predicts the water quality change trend, such as key indexes of total phosphorus, ammonia nitrogen, chemical oxygen demand and the like, and is important for water resource management, water pollution control and sudden water environment event early warning. The traditional mechanism model depends on a complex physicochemical equation, and has the problems of difficult parameter calibration, high calculation cost and the like. In recent years, a deep learning time sequence prediction model represented by a cyclic neural network and a transducer has great potential in the field of water quality prediction due to strong nonlinear fitting capability. However, existing neural network-based water quality prediction methods still have significant limitations. Firstly, the internal structure and the induced deviation of different neural network models are different, so that the capture capability of the neural network models on different scales and different mode characteristics in water quality time sequence data is different. It is difficult for a single model to maintain optimal performance in all scenarios, and there are instability and contingency in the predicted results. Second, more critical, existing methods mostly only provide deterministic point predictions, but fail to quantitatively evaluate the uncertainty of the predictions. In practical applications, the decision maker needs to know not only what the predicted value is, but also the possible fluctuation range of the predicted value and the confidence level thereof, namely the uncertainty of the prediction. The lack of uncertainty information severely limits the reliability and decision support value of the predicted outcome. Although research attempts have been made to integrate multiple prediction models by means of model averaging or simple weighting, such methods generally rely on fixed weights or simple linear combinations, and still stay at the point prediction level, failing to describe complex dependencies between multiple model predictions and their actual observations from a probability perspective, so that probability prediction and uncertainty quantization cannot be achieved in a true sense. Disclosure of Invention The invention provides a water quality prediction method, device, equipment and medium integrating uncertainty quantification, which are used for solving the problems that single model prediction is unstable and uncertainty of prediction cannot be quantified in the prior art. The invention provides a water quality prediction method integrating uncertainty quantification, which comprises the steps of obtaining environment driving factor values at future time, inputting the environment driving factor values into a plurality of pre-built water quality prediction models to obtain future time water quality prediction values output by the water quality prediction models, respectively training a plurality of network models by a historical data set to obtain the pre-built water quality prediction models, wherein the historical data set comprises a historical water quality observation sequence and a historical environment driving factor sequence, the historical water quality observation sequence comprises water quality observation values at a plurality of historical time, the historical environment driving factor sequence comprises environment driving factor values at a plurality of historical time, and generating a prediction interval and a final future time water quality prediction reference value according to the future time water quality prediction values output by the water quality prediction models and a pre-built joint probability distribution model, wherein the joint probability distribution model is used for describing statistical dependency relationship between the prediction values and the observation values. According to the method, a plurality of network models are integrated, the differentiation capability of each model in the aspects of capturing time sequence dependence and local characteristic and long-range association is comprehensively utilized, the structural deviation and the contingency of a single model are effectively overcome, the final prediction result is more accurate and stable, the excessive dependence of the prediction result on a specific model or data characteristic is reduced, and the overall robustness of the method is improved. And then, introducing a joint probability distribution model, so that the conditional probability distribution of the real water quality value under the given multi-model prediction information, namely a prediction interval and a fin